A forecast on this site is a dated, public claim that a topic will see a spike in domain registrations inside a stated window. It is published before the outcome is known and graded afterwards, which means the cost of publishing a bad one is permanent. The engine is built around that cost, and most of what follows is about the conditions under which it declines to say anything at all.
Four conditions, all of which must hold
A topic is only eligible for a first forecast when every one of these is true. They are fixed conditions in the engine, not judgement calls made per topic, and the exact levels are not published.
- Several distinct kinds of evidenceOne source moving is a source moving. Several independent kinds moving together is a topic moving.
- A baseline for eachEnough readings to measure a rise against, rather than a number and an opinion about it.
- The weekly cycle removedEnough history to divide it out, so a window with one fewer weekend does not look like movement.
- Above the engine's own floorBelow it the evidence points nowhere in particular, so the engine says nothing.
A topic that fails any of these is watched, not forecast, and its page says so in those words rather than showing a zero. At launch that is the normal state for nearly every topic on the trend board, which is what a young record honestly looks like.
What the probability actually is
The published probability is a function of the composite score and nothing else. It is deliberately bounded well away from certainty, because a registration spike depends on human attention that no set of leading indicators fully determines.
So the strongest reading this engine can take still produces a claim. A published probability is a calibration target, and the calibration curve on the track record is how you check whether the target is being met.
The lead window, and why a stronger signal gets a shorter one
Every forecast carries a window in days rather than a single date. A higher score produces a narrower and nearer window, because the point of the signal mix is that it leads the crowd, and a stronger reading means the crowd is closer behind. The far edge of the window becomes the evaluation date, so this choice also decides when the engine has to face the result.
The threshold is fixed before the outcome
What counts as a hit is set at the moment the forecast is published, stored with it, and never recalculated. It is a multiple of that topic’s own observed baseline rather than an absolute count, so it means the same thing for a large topic and a small one.
The reason to store it rather than compute it at evaluation is the whole reason this section exists. A threshold chosen after the fact is not a forecast, it is a description, and a track record built from descriptions is worth nothing. This is also why the evaluation code reads the stored threshold and has no access to the function that would produce a new one.
Where to look next
How a forecast is graded covers what happens at the evaluation date, including the outcome that is neither a hit nor a miss. If you want to see the mechanism applied rather than described, the published record carries one page per graded call.